arrow_backBack to the overview
The engine

The part nobody should build twice.

Objectives that hold. Records you can trust. Models you can swap.

The engine is where the difficult, general work lives. It assembles context before a conversation starts, holds working memory during it, enforces conversation policy turn by turn, executes the objective ladder, mediates every tool call, and produces a structured record at the end. It is channel agnostic and language agnostic by construction, so the same worker logic drives a phone call, a message thread and a web chat in any supported language.

Proceduralobjective ladders that cannot be skipped
Evidence boundrecords populated only from what was actually said
Configurablemodels, thinking budgets and voice set at setup
Primary objectives

An ordered ladder, not a goal.

Most conversational agents are given a goal. Hub workers are given an ordered, procedural sequence in which each rung is a materially different outcome, and a lower rung may only be attempted once the higher one is genuinely closed off.

100
The primary meetingThe first substantive ask on every call, before any other meeting is discussed. Work through the contact or their assistant to reach the calendar.
80
An alternate executive or direct reportOnly once the primary is declined, unavailable or redirected. Agree a day and time first, then ask once whether the primary might still join. Never stacked into a single question.
55
Account team follow upIf nobody suitable can meet inside the window, a follow up with the local team within thirty days.
30
CallbackA distant fourth. A tentative meeting beats a callback, and a callback beats nothing at all.
shield
Anti-degradation is built in. A contact saying the executive "may join" is recorded as the lower rung achieved plus a partial upgrade in play, never as the higher rung claimed. Once a contact says they will check a calendar, the topic is settled permanently and the worker will not ask again.

Standing objectives run alongside

Intelligence captured opportunistically where the conversation allows, never worked as a checklist and never at the cost of the ladder. They are the difference between a booking service and an intelligence asset.

sentimentwillingness to engageinitiatives + funding stageprimary pain who else shares itreferral + level securedcompetitors + contract timing new hires and new roleschannel consentorganisations mentionedcorrections to our assumptions
low_priority
Precedence is enforced by the engine, not the prompt. Respect for the person, covering consent, honesty and easy exits, outranks brevity, which outranks the objective ladder, which outranks bonus intelligence. Every worker inherits that stack.
The outcome contract

Every interaction ends in a record, not a summary.

A narrative summary cannot be reported on, compared, or scored. A structured contract can. This is the object every downstream system reads from.

Field groupWhat it holds
OutcomeA required level from a closed set matching the ladder, a one line summary of what was agreed, and a status distinguishing confirmed, where a day and time were explicit, from tentative
Upgrade & confirmationA higher outcome partially in play but not confirmed, and how and when anything tentative will be confirmed, as agreed with the contact
DispositionSentiment toward your company, and readiness to engage further
Commercial intelligenceInitiatives with funding stage and timing, the principal pain, who else shares the concern, competitive presence and contract timing, new leaders and newly created roles
RelationshipReferrals with the level secured, being a name, permission to reference, or a promised introduction; channel permissions and stated preferences; every organisation and person mentioned with a clause of context
CorrectionsWhich of our prior assumptions were confirmed or corrected, with the corrected fact
SuppressionAny explicit request for no further contact, recorded in the contact's own terms with its scope

Fields are populated only from evidence in the conversation, and anything unknown is omitted rather than invented. A record built on inference would be worthless the moment anyone relied on it.

Prompt architecture

Six versioned layers, not one prompt.

A single concatenated prompt cannot be safely edited by a customer, cannot be improved centrally without touching every deployment, and cannot tell you which change caused a regression. Layers solve all three.

Layer 1 · closedKernel

The invariant core: the priority stack, honesty and disclosure rules, the settled topic ledger, turn discipline, ladder mechanics, tool and contract semantics, the ending protocol. Corvair engineering only.

Layer 2Agent template

Personality, register, voice binding, gendered language handling, conversational style defaults. Selected by you from the template library.

Layer 3Application

The domain: inside sales, property, collections. Vocabulary, outcome level names, intelligence field labels, domain guardrails.

Layer 4 · yoursClient

Company identity, brand voice, product and pricing boundaries, escalation contacts, jurisdictional posture. Edited by your administrator, within limits the kernel enforces.

Layer 5Scenario

The engagement briefing and its objective ladder, authored by your sales operations team.

Layer 6Session

Contact, account intelligence, memory, prior interaction notes, language and channel state. Assembled by the runtime and never hand edited.

lock
The kernel is never edited to fix an application problem, and a client layer can never weaken a kernel rule. That is what makes it safe to let your administrator tune layer four, and what lets a kernel improvement reach every worker at once without touching a single customer configuration. Every interaction records the exact layer versions in force, so when a score moves, the responsible layer is identifiable rather than inferred.
Models, thinking and voice

Nothing assumes a model, a voice, or a reasoning budget.

Conversational latency and analytical depth are opposing requirements, and they should never be served by one configuration. Each worker carries an engine profile binding a model class to each task class it performs.

Task classRequirementTypical binding
Live conversationSub second first token, streaming audio, interruption handling. Depth traded for latencyA low latency live model
In conversation retrievalFast, grounded, cited. Runs inside a turn without stalling itA fast model against the substrate
Post conversation analysisRigorous, evidence bound, schema conformant. Latency irrelevantA deliberative reasoning model
Coaching and revisionThe highest reasoning demand in the system. Must produce surgical, defensible editsA deliberative model, elevated thinking
Research and briefingLong running, tool using, grounded in searchA deliberative model with search grounding
Extraction and enrichmentHigh volume, structured, cheapA small fast model

Voice, behind an interface

Google is the first voice provider and ElevenLabs is the named alternate. Two implementations rather than one is what makes the interface real rather than aspirational, because an interface with a single implementation behind it is a guess.

  • Synthesis, streaming and barge-in detection
  • Voicemail detection with a configured strategy
  • Voice profiles curated per language, since a voice that sounds warm in English can sound clipped elsewhere
  • Latency reporting per provider, which is how the two get compared honestly

Thinking level, per task class

A live turn wants minimal deliberation. An evaluator scoring a transcript wants a great deal. Fixing one value platform wide either makes conversations sluggish or makes analysis shallow.

  • Engine, thinking level, token and latency budget, and a fallback binding per task class
  • Bound during setup or from the administration screen, never in code
  • Priced and logged per interaction, so a conversation has a known cost
  • Changing a binding is configuration, not a release
Language

Following the contact is a concept, not a flag.

Sixteen languages today, with detection and following provided as a runtime service rather than as agent logic.

translate

Follow the contact

Detect and continue in their language, including mid conversation switches, with correct gendered grammar and politeness where the language requires it.

record_voice_over

Transcribe verbatim

The live transcript is never translated or rewritten. It is the evidentiary record of what was actually said.

sync_alt

Translate for the record

A parallel rendering in your record language, so structured fields, notes and CRM evidence stay consistent regardless of what was spoken.

display_settings

Interface language apart

The language a contact is served in, the language records are kept in, and the language your team reads dashboards in are three separate settings.

inventory_2
Raw and translated are both retained, always. The verbatim transcript is the evidence and the translation is the working artefact. A dispute, an audit or a quality review may need either, so discarding one destroys the value of everything built on it.
Memory

Three tiers, over a canonical record.

When a conversation moves from a message thread to a voice call, the context travels with the contact, so the worker on the call knows exactly what was discussed in text ten minutes earlier.

One interaction

Working memory

The current thread, the contact's recent statements, the live objective rung, the settled topic ledger and the next step options.

30 to 90 days

Short term memory

Recent conversation summaries, pending actions, referral chains in progress, qualification captured, and the state of each active flow.

Permanent

Long term memory

Relationship history, stakeholder profiles, initiative signals, competitive intelligence. Institutional memory that survives personnel change.

Derived

Corvair Memory

A canonical, attributed chronology of decisions, actions and pivots, from which working context is derived rather than stored, so continuity stays explainable.

psychology_alt
Account intelligence is treated as hypothesis, not fact. Workers reference it naturally and invite correction, because corrections are wins. When something captured contradicts a prior assumption, the discrepancy is flagged, the record updated, the confidence adjusted and the change surfaced to your team.
Next

The engine is only half the story.

Everything that varies by customer, channel or system sits behind an adapter. That is what lets one runtime serve a seller, a collections specialist and a property assistant without a fork.